Unlike traditional models that use physical properties in a location to simulate physical processes, machine-learning models are largely black boxes that train on past patterns and spit out predictions of future patterns. But Weather Next 3 is adding a tiny bit of physical information to calculate surface temperature and dew point at any specific location you want to pull up. It checks whether that point is land or ocean and uses its surface elevation. By training on past weather station data tagged with that information, the team says they get better forecast predictions.

Some oddities: The white paper shows some results to document forecast performance improvements over Weather Next 2, as well as the European Centre for Medium-Range Weather Forecasts (ECMWF) AI model. They note a roughly 5 percent improvement in upper atmosphere condition accuracy over their previous model, for example, which they say equates to about six more hours of accurate forecast lead time. And their change to calculating surface temperature for a specific location improved accuracy by up to 30 percent. They’re generally beating the ECWMF model on these metrics as well.

There is one curious exception that the paper doesn’t even guess at the cause of. For a number of variables, their comparison to the initial six-hours-ahead forecast from the other models shows Weather Next 3 doing worse before pulling ahead for the rest of a 15-day forecast. Larger-scale patterns are also not without some weirdness. You can see the shape of the model’s grid in some predictions, like the map of precipitation showing some distinctly hexagonal blobs.

Overall, the team says their new model “represents a major step forward for AI-based weather predictions by going beyond relying purely on analysis and utilizing information-dense, low-latency observation data.” Weather Next 3 is now the source of forecast information across Google services, including Search, Gemini, and Maps.